Long-horizon events forecasting is a crucial task across various domains, including retail, finance, healthcare, and social networks. Traditional models for event sequences often extend to forecasting on horizon using an autoregressive (recursive) multi-step strategy, which has limited effectiveness due to typical convergence to constant or repetitive outputs. To address this limitation, we introduce DEF, a novel approach for simultaneous forecasting of multiple future events on a horizon with high accuracy and diversity. Our method optimally aligns predictions with ground truth events during training by using a novel matching-based loss function. We establish a new state-of-the-art in long-horizon event prediction, achieving up to a 50% relative improvement over existing temporal point processes and event prediction models. Furthermore, we achieve state-of-the-art performance in next-event prediction tasks while demonstrating high computational efficiency during inference.

Accurate forecasting of multivariate time series data remains a formidable challenge, particularly due to the growing complexity of temporal dependencies in real-world scenarios. While neural network-based models have achieved notable success in this domain, complex channel-dependent models often suffer from performance degradation compared to channel-independent models that do not consider the relationship between components but provide high robustness due to small capacity. In this work, we propose HN-MVTS, a novel architecture that integrates the hypernetwork-based generative prior with an arbitrary neural network forecasting model. The input of this hypernetwork is a learnable embedding matrix of time series components. To restrict the number of new parameters, the hypernetwork learns to generate the weights of the last layer of the target forecasting networks, serving as a data-adaptive regularizer that improves generalization and long-range predictive accuracy. The hypernetwork is only used during training, so it does not increase the inference time compared to the base forecasting model. Extensive experiments on eight benchmark datasets demonstrate that application of HN-MVTS to the state-of-the-art models (DLinear, PatchTST, TSMixer, etc.) typically improves their performance. Our findings suggest that hypernetwork-driven parameterization offers a promising direction for enhancing existing forecasting techniques in complex scenarios.

Alexandra Bazarova, Andrei Volodichev, Aleksandr Yugay, Andrey Shulga, Alina Ermilova, Konstantin Polev, Julia Belikova, Rauf Parchiev, Dmitry Simakov, Maxim Savchenko, Andrey Savchenko, Serguei Barannikov, Alexey Zaytsev. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Andrey Veprikov, Vladimir Solodkin, Zyl Alexander, Andrey Savchenko, Aleksandr Beznosikov. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Nowadays, the success of financial organizations heavily depends on their ability to process digital traces generated by their clients, e.g., transaction histories, gathered from various sources to improve user modeling pipelines. As general-purpose LLMs struggle with time-distributed tabular data, production stacks still depend on specialized tabular and sequence models that are limited in transferability and require labeled data. To address this, we introduce FinTRACE, a retrieval-first architecture that converts raw transactions into reusable feature representations, applies rule-based detectors, and stores the resulting signals in a behavioral knowledge base with graded associations to the objectives of downstream tasks. Across public and industrial benchmarks, FinTRACE substantially improves low-supervision transaction analytics, doubling zero-shot MCC on churn prediction performance from 0.19 to 0.38 and improving 16-shot MCC from 0.25 to 0.40. We further use FinTRACE to ground LLMs via instruction tuning on retrieved behavioral patterns, achieving state-of-the-art LLM results on transaction analytics problems.

Financial organizations collect a huge amount of temporal (sequential) data about clients, which is typically collected from multiple sources (modalities). Despite the urgent practical need, developing deep learning techniques suitable to handle such data is limited by the absence of large open-source multi-source real-world datasets of event sequences. To fill this gap, which is mainly caused by security reasons, we present the first industrial-scale publicly available multimodal banking dataset, MBD, that contains information on more than 2M corporate clients of a large bank. Clients are represented by several data sources: 950M bank transactions, 1B geo position events, 5M embeddings of dialogues with technical support, and monthly aggregated purchases of four bank products. All entries are properly anonymized from real proprietary bank data, and the experiments confirm that our anonymization still saves all significant information for introduced downstream tasks. MBD enables supports campaigning task (predict future customer purchases). We provide numerical results for the state-of-the-art event sequence modeling techniques demonstrate the superiority of fusion baselines over single-modal techniques for this task. HuggingFace Link: https://huggingface.co/datasets/ai-lab/MBD Github Link: https://github.com/Dzhambo/MBD

Though Large Vision-Language Models (LVLMs) are being actively explored in medicine, their ability to conduct complex realworld telemedicine consultations combining accurate diagnosis with professional dialogue remains underexplored.This paper presents 3MDBench (Medical Multimodal Multi-agent Dialogue Benchmark), an open-source framework for simulating and evaluating LVLM-driven telemedical consultations.3MDBench simulates patient variability through temperament-based Patient Agent and evaluates diagnostic accuracy and dialogue quality via Assessor Agent.It includes 2996 cases across 34 diagnoses from real-world telemedicine interactions, combining textual and image-based data.The experimental study compares diagnostic strategies for widely used open and closed-source LVLMs.We demonstrate that multimodal dialogue with internal reasoning improves F1 score by 6.5% over non-dialogue settings, highlighting the importance of context-aware, informationseeking questioning.Moreover, injecting predictions from a diagnostic convolutional neural network into the LVLM's context boosts F1 by up to 20%.

While current time series research focuses on developing new models, crucial questions of selecting an optimal approach for training such models are underexplored. Tsururu, a Python library introduced in this paper, bridges SoTA research and industry by enabling flexible combinations of global and multivariate approaches and multi-step-ahead forecasting strategies. It also enables seamless integration with various forecasting models. Available at https://github.com/sb-ai-lab/tsururu.

The domain of event sequences is widely applied in various industrial tasks in banking, healthcare, etc., where temporal tabular data processing is required. This paper introduces PyTorch-Lifestream, the first open-source library specially designed to handle event sequences. It supports scenarios with multimodal data and offers a variety of techniques for learning embeddings of event sequences and end-to-end model training. Furthermore, PyTorch-Lifestream efficiently implements state-of-the-art methods for event sequence analysis and adapts approaches from similar domains, thus enhancing the versatility and performance of sequence-based models for a wide range of applications, including financial risk scoring, campaigning, user ID matching, churn prediction, fraud detection, medical diagnostics, and recommender systems.

Progress in remote PhotoPlethysmoGraphy (rPPG) is limited by the critical issues of existing publicly available datasets: small size, privacy concerns with facial videos, and lack of diversity in conditions. The paper introduces a novel comprehensive large-scale multi-view video dataset for rPPG and health biomarkers estimation. Our dataset comprises 3600 synchronized video recordings from 600 subjects, captured under varied conditions (resting and post-exercise) using multiple consumer-grade cameras at different angles. To enable multimodal analysis of physiological states, each recording is paired with a 100 Hz PPG signal and extended health metrics, such as electrocardiogram, arterial blood pressure, biomarkers, temperature, oxygen saturation, respiratory rate, and stress level. Using this data, we train an efficient rPPG model and compare its quality with existing approaches in cross-dataset scenarios. The public release of our dataset and model should significantly speed up the progress in the development of AI medical assistants.

Since their introduction, Transformer-based models, such as SASRec and BERT4Rec, have become common baselines for sequential recommendations, surpassing earlier neural and non-neural methods. A number of following publications have shown that the effectiveness of these models can be improved by, for example, slightly updating the architecture of the Transformer layers, using better training objectives, and employing improved loss functions. However, the additivity of these modular improvements has not been systematically benchmarked - this is the gap we aim to close in this paper. Through our experiments, we identify a very strong model that uses SASRec's training objective, LiGR Transformer layers, and Sampled Softmax Loss. We call this combination eSASRec (Enhanced SASRec). While we primarily focus on realistic, production-like evaluation, in our preliminarily study we find that common academic benchmarks show eSASRec to be 23% more effective compared to the most recent state-of-the-art models, such as ActionPiece. In our main production-like benchmark, eSASRec resides on the Pareto frontier in terms of the accuracy-coverage tradeoff (alongside the recent industrial models HSTU and FuXi. As the modifications compared to the original SASRec are relatively straightforward and no extra features are needed (such as timestamps in HSTU), we believe that eSASRec can be easily integrated into existing recommendation pipelines and can can serve as a strong yet very simple baseline for emerging complicated algorithms. To facilitate this, we provide the open-source implementations for our models and benchmarks in repository https://github.com/blondered/transformer_benchmark

In this paper, we introduce a novel tool for speech emotion recognition, CA-SER, that borrows self-supervised learning to extract semantic speech representations from a pre-trained wav2vec 2.0 model and combine them with spectral audio features to improve speech emotion recognition. Our approach involves a self-attention encoder on MFCC features to capture meaningful patterns in audio sequences. These MFCC features are combined with high-level representations using a multi-head cross-attention mechanism. Evaluation of speech emotion recognition on the IEMOCAP dataset shows that our system achieves a weighted accuracy of 74.6%, outperforming most existing techniques.

This paper introduces InsideOut, an original innovative framework that augments the emotional intelligence of Large Language Models (LLMs). Motivated by the cartoon, InsideOut is designed around a net of specialized agents, each dedicated to one of Ekman’s fundamental emotions. These agents collaboratively refine responses sensitive to the emotional context of interactions. Our assessments, conducted using EmpatheticDialogues and involving models like GPT-4 and GigaChat, indicate substantial improvements in identifying human emotions and generating empathetic responses. These improvements are most evident in situations with apparent valence-arousal differences. InsideOut offers a promising avenue for evolving AI into more perceptive and human-centric communicators.

We address the challenge of devising neural network architectures to extract facial descriptors across diverse mobile and edge devices. Employing neural architecture search, we introduce a novel framework that selects optimal subnetworks from a SuperNet using an evolutionary search. Using a surrogate gradient boosting classifier to avoid direct accuracy estimation of subnetworks on validation sets, our approach swiftly delivers the most efficient and accurate models tailored to specific devices within minutes. Demonstrating versatility through an Android demo app, our framework excels in tasks like face recognition and emotion understanding across various devices, achieving real-time processing and superior accuracy compared to existing mobile models.

As machine learning increasingly becomes a crucial tool for geospatial data analysis, finding and deploying a suitable model presents significant challenges, including the need for expertise in both programming and geospatial analysis, organizing data flow, and accurately assessing the results. To address these challenges, this paper introduces GlobeFlowGPT, a multimodal, chat-based framework designed to meet these demands by integrating domain-specific tools, machine learning models, Multimodal Large Language Models, and essential operational data. It leverages a Large Language Model orchestrator, facilitating complex geospatial tasks through a conversational interface. GlobeFlowGPT's flexible, containerized architecture allows for the rapid integration of cutting-edge models tailored for geospatial data, ensuring that the framework remains scalable and relevant amid ongoing technological advancements. We demonstrate the ability of our framework to streamline the analysis of geospatial data and expand the capabilities of modern MLLMs with complex geospatial machine learning models.

In the rapidly evolving domain of Recommender Systems (RecSys), new algorithms frequently claim state-of-the-art performance based on evaluations over a limited set of arbitrarily selected datasets. However, this approach may fail to holistically reflect their effectiveness due to the significant impact of dataset characteristics on algorithm performance. Addressing this deficiency, this paper introduces a novel benchmarking methodology to facilitate a fair and robust comparison of RecSys algorithms, thereby advancing evaluation practices. By utilizing a diverse set of 30 open datasets, including two introduced in this work, and evaluating 11 collaborative filtering algorithms across 9 metrics, we critically examine the influence of dataset characteristics on algorithm performance. We further investigate the feasibility of aggregating outcomes from multiple datasets into a unified ranking. Through rigorous experimental analysis, we validate the reliability of our methodology under the variability of datasets, offering a benchmarking strategy that balances quality and computational demands. This methodology enables a fair yet effective means of evaluating RecSys algorithms, providing valuable guidance for future research endeavors.

One of the urgent tasks of artificial intelligence is to assess the safety and alignment of large language models (LLMs) with human behavior. Conventional verification only in pure natural language processing benchmarks can be insufficient. Since emotions often influence human decisions, this paper examines LLM alignment in complex strategic and ethical environments, providing an in-depth analysis of the drawbacks of our psychology and the emotional impact on decision-making in humans and LLMs. We introduce the novel EAI framework for integrating emotion modeling into LLMs to examine the emotional impact on ethics and LLM-based decision-making in various strategic games, including bargaining and repeated games. Our experimental study with various LLMs demonstrated that emotions can significantly alter the ethical decision-making landscape of LLMs, highlighting the need for robust mechanisms to ensure consistent ethical standards. Our game-theoretic analysis revealed that LLMs are susceptible to emotional biases influenced by model size, alignment strategies, and primary pretraining language. Notably, these biases often diverge from typical human emotional responses, occasionally leading to unexpected drops in cooperation rates, even under positive emotional influence. Such behavior complicates the alignment of multiagent systems, emphasizing the need for benchmarks that can rigorously evaluate the degree of emotional alignment. Our framework provides a foundational basis for developing such benchmarks.

We develop and evaluate neural architectures to model the user behavior in recommender systems (RS) inspired by click models for Web search but going beyond standard click models. Proposed architectures include recurrent networks, Transformer-based models that alleviate the quadratic complexity of self-attention, adversarial and hierarchical architectures. Our models outperform baselines on the ContentWise and RL4RS datasets and can be used in RS simulators to model user response for RS evaluation and pretraining.

In this paper, we consider the problem of the high computational complexity of video-based facial expression recognition. A novel sequential procedure is proposed with an adaptive frame rate selection in a short video fragment to speed up decision-making. We automatically adjust the frame rate and process fewer frames with a low frame rate for more straightforward videos and more frames for complex ones. To determine the frame rate at which an inference is sufficiently reliable, the Benjamini-Hochberg procedure from multiple comparisons theory is employed to control the false discovery rate. The main advantages of our method are an improvement of the trustworthiness of decision-making by maintaining only one hyper-parameter (false acceptance rate) and its applicability with arbitrary neural network models used as facial feature extractors without the need to re-train these models. An experimental study on datasets from ABAW and EmotiW challenges proves the superior performance (1.5-40 times faster) of the proposed approach compared to processing all frames and existing techniques with early exiting and adaptive frame selection.

Recommender systems are essential for personalized content delivery and have become increasingly popular recently. However, traditional recommender systems are limited in their ability to capture complex relationships between users and items. Dynamic graph neural networks (DGNNs) have recently emerged as a promising solution for improving recommender systems by incorporating temporal and sequential information in dynamic graphs. In this paper, we propose a novel method, "Ti-DC-GNN" (Time-Interval Dual Causal Graph Neural Networks), based on an intermediate representation of graph evolution as a sequence of time-interval graphs. The main parts of the method are the novel forms of interval graphs: graph of causality and graph of consequence that explicitly preserve inter-relationships between edges (user-items interactions). The local and global message passing are developed based on edge memory to identify short-term and long-term dependencies. Experiments on several well-known datasets show that our method consistently outperforms modern temporal GNNs with node memory alone in dynamic edge prediction tasks.

One of the main issues in face identification is to create a real-time application with high accuracy. Images are presented by high-dimensional feature vectors that are produced by convolutional neural networks. In order to effectively process such vectors, the hierarchical algorithm was proposed in this paper that applies sequential analysis to search the nearest neighbors among the reference images of the most reliable classes selected at a certain algorithm level. Principal component analysis was also applied to select the most significant part of feature vectors. Moreover, another issue of face recognition was investigated: the lack of training data of specific types (bad quality image, different scale or illumination, children/old people, etc.), The recognition accuracy may be low for input images that are not similar to the majority of images in the dataset used to train the feature extractors. Therefore, we propose the level of preprocessing input images by detection of rare data. In this paper datasets such as VGGFace2, MS-Celeb-1M, All ages faces, Large age gap face, and different facial descriptors were used for the testing goal. Also, we provide a procedure to collect a special dataset for a given training set by using different transformations and automatic detection of anomalies.

In this article, behaviour of students in the e-learning environment is analyzed. The novel pipeline is proposed based on video facial processing. At first, face detection, tracking and clustering techniques are applied to extract the sequences of faces of each student. Next, a single efficient neural network is used to extract emotional features in each frame. This network is pre-trained on face identification and fine-tuned for facial expression recognition on static images from AffectNet using a specially developed robust optimization technique. It is shown that the resulting facial features can be used for fast simultaneous prediction of students’ engagement levels (from disengaged to highly engaged), individual emotions (happy, sad, etc.,) and group-level affect (positive, neutral or negative). This model can be used for real-time video processing even on a mobile device of each student without the need for sending their facial video to the remote server or teacher's PC. In addition, the possibility to prepare a summary of a lesson is demonstrated by saving short clips of different emotions and engagement of all students. The experimental study on the datasets from EmotiW (Emotion Recognition in the Wild) challenges showed that the proposed network significantly outperforms existing single models.

The article considers the possibility to predict preferences of a user given a set of his or her photos. The novel algorithm is introduced with detection of private photos based on facial clustering. Such photos are analyzed on the mobile device with the lightweight neural networks. The remaining photos may be sent to the remote server to be processed by the state-of-the-art models. In both cases each photo is associated with the outputs of neural networks trained on scene classification and object detection tasks. Qualitative and quantitative results demonstrate the potential of the proposed approach in event recognition and understanding of preferences of celebrities given their real photos from social networks.

Understanding image advertisements is a challenging task, often requiring non-literal interpretation.We argue that standard image-based predictions are insufficient for symbolism prediction.Following the intuition that texts and images are complementary in advertising, we introduce a multimodal ensemble of a state of the art image-based classifier, a classifier based on an object detection architecture, and a fine-tuned language model applied to texts extracted from ads by OCR.The resulting system establishes a new state of the art in symbolism prediction.

In this paper, we consider the problem of event recognition on single images. In contrast to conventional fine-tuning of convolutional neural networks (CNN), we proposed to use image captioning, i.e., a generative model that converts images to textual descriptions. The motivation here is the possibility to combine conventional CNNs with a completely different approach in an ensemble with high diversity. As event recognition task has nothing serial or temporal, obtained captions are one-hot encoded and summarized into a sparse feature vector suitable for the learning of an arbitrary classifier. We provide the experimental study of several feature extractors for Photo Event Collection, Web Image Dataset for Event Recognition and Multi-Label Curation of Flickr Events Dataset. It is shown that the image captions trained on the Conceptual Captions dataset can be classified more accurately than the features from an object detector, though they both are obviously not as rich as the CNN-based features. However, an ensemble of CNN and our approach provides state-of-the-art results for several event datasets.

In this paper the problem of high computational complexity of deep convolutional nets in image recognition is considered. An existing framework of adaptive neural networks is extended by appending the separate classifier to intermediate layers. The hierarchical representations of the input image are sequentially analyzed. If the first classifier returns rather high confidence score, the inference process will be terminated. Otherwise, the inference to the next intermediate layer with attached classifier is continued until the reliable solution is obtained or the penultimate layer is reached. The thresholds for classifier scores at each layer are automatically chosen based on the Benjamini-Hochberg multiple comparisons for a specified confidence level. Experimental study for both pre-trained and fine-tuned deep convolutional neural networks demonstrates that the proposed approach reduces the running time by up to 1.7 times without significant accuracy degradation. Moreover, the larger is the training sample, the more noticeable is the gain in performance.

In this paper a new formulation of event recognition task is examined: it is required to predict event categories given a gallery of images, for which albums (groups of photos corresponding to a single event) are unknown. The novel two-stage approach is proposed. At first, features are extracted in each photo using the pre-trained convolutional neural network (CNN). These features are classified individually. The normalized scores of the classifier are used to group sequential photos into several clusters. Finally, the features of photos in each group are aggregated into a single descriptor using neural attention mechanism. This algorithm is implemented in Android mobile application. Experimental study with features extracted by contemporary convolutional neural networks including EfficientNets for Photo Event Collection and Multi-Label Curation of Flickr Events Dataset demonstrates that the proposed approach is 9-23% more accurate than conventional event recognition on single photos. Moreover, proposed method has 13-16% lower error rate when compared to classification of groups of photos obtained with hierarchical clustering of CNN-based embeddings.

If the training data set in image recognition task is not very large, the feature extraction with a convolutional neural network is usually applied. Here, we focus on the nonparametric classification of extracted feature vectors using the probabilistic neural network (PNN). The latter is characterized by the high runtime and memory space complexity. We propose to overcome these drawbacks by replacing the exponential activation function in the Gaussian kernel to the complex exponential functions. Such complex nonlinearities make it possible to accurately approximate the unknown density function using the network with the number of neurons proportional to only cubic root of the database size. As a result, the proposed approach decreases the runtime and memory complexities of the PNN without losing its main advantages, namely, fast training and convergence to the Bayesian decision. In the experimental study, we describe a protocol for comparing recognition methods using the well-known visual object category data sets in the context of the small sample size problem. It has been experimentally shown that our approach rapidly obtains accurate decisions when compared to the known classifiers including the baseline PNN.

In this paper we deal with unconstrained face recognition with few training samples. The facial images are described with the off-the shelf high-dimensional features extracted with a deep convolutional neural network (CNN), which was preliminarily trained with an external very-large dataset. We focus on drawbacks of conventional probabilistic neural network (PNN), namely, low recognition performance and high memory space complexity. We propose to modify the PNN by replacing the exponential activation function in the Gaussian Parzen kernel to the trigonometric functions and use the orthogonal series density estimation of the CNN features. We demonstrate that the proposed approach significantly decreases the runtime complexity of face recognition if the classes are rather balanced and there are more than five training images per each subject. An experimental study with either traditional VGGNet and Light CNN, or contemporary VGGFace2_ft and MobileNet trained on VGGFace-2 dataset, have shown that our algorithm is very efficient and rather accurate in comparison with the instance-based learning classifiers.

In this paper, we describe our algorithmic approach, which was used for submissions in the fifth Emotion Recognition in the Wild (EmotiW 2017) group-level emotion recognition sub-challenge. We extracted feature vectors of detected faces using the Convolutional Neural Network trained for face identification task, rather than traditional pre-training on emotion recognition problems. In the final pipeline an ensemble of Random Forest classifiers was learned to predict emotion score using available training set. In case when the faces have not been detected, one member of our ensemble extracts features from the whole image. During our experimental study, the proposed approach showed the lowest error rate when compared to other explored techniques. In particular, we achieved 75.4% accuracy on the validation data, which is 20% higher than the handcrafted feature-based baseline. The source code using Keras framework is publicly available.